I have a benchmark (process A) which launches another process (process B)
and the two send messages back and forth through a couple of sockets, one
for sending and the other for receiving. Both the processes have two
go-routines for handling the sending and the receiving side.
Process A sends a message to B using the "encoding/gob" package to write to
the socket which is subsequently decoded. After sending a message, A waits
for a response from B. B does some processing on the message and sends a
reply back.
The round trip time of getting a response time seemed to be highly
variable. I found that the runtime is taking up the bulk of the running
time. Also after receiving a message from B, the waiting go-routine takes a
long time to wake up, even taking 0.5x to 1x time as the actual round trip
time.
Initially the round trip time is around 200 microseconds and it drops to 20
microseconds as more messages are sent. The runtime seems to "learn" the
best way to minimize the running time eventually.
How can I reduce the initial overhead incurred by the go runtime? Is there
a way to enforce or tweak the scheduling given the above scenario?
Benchmark info from pprof:
flat flat% sum% cum cum%
93.87ms 33.56% 33.56% 98.80ms 35.33% syscall.Syscall
79.91ms 28.57% 62.14% 79.91ms 28.57% runtime.futex
7.15ms 2.56% 64.69% 54.69ms 19.55% runtime.findrunnable
6.30ms 2.25% 66.94% 6.62ms 2.37% runtime.runqgrab
3.22ms 1.15% 68.10% 3.22ms 1.15% runtime.(*randomEnum).next
2.68ms 0.96% 69.05% 3.96ms 1.42% runtime.deferreturn
2.32ms 0.83% 69.88% 2.38ms 0.85% runtime.lock
2.30ms 0.82% 70.71% 2.30ms 0.82% runtime.memmove
2.24ms 0.8% 71.51% 5.77ms 2.06% runtime.mallocgc
2.21ms 0.79% 72.30% 2.22ms 0.79% runtime.unlock
2.02ms 0.72% 73.02% 2.25ms 0.8% time.now
2.01ms 0.72% 73.74% 8.63ms 3.09% runtime.runqsteal
1.78ms 0.64% 74.37% 1.80ms 0.64% runtime.casgstatus
1.63ms 0.58% 74.96% 3.38ms 1.21% runtime.exitsyscall
1.60ms 0.57% 75.53% 33.05ms 11.82% runtime.stopm
1.56ms 0.56% 76.09% 3.16ms 1.13% runtime.selectgo
1.53ms 0.55% 76.63% 29.86ms 10.68% runtime.notesleep
1.38ms 0.49% 77.13% 1.41ms 0.5% runtime.newdefer
1.30ms 0.46% 77.59% 7.74ms 2.77% fmt.(*pp).doPrintf
1.17ms 0.42% 78.01% 1.95ms 0.7% runtime.mapaccess2
1.03ms 0.37% 78.38% 75.87ms 27.13% runtime.schedule
1ms 0.36% 78.74% 53.46ms 19.11% runtime.notewakeup
0.83ms 0.3% 79.03% 2.34ms 0.84% runtime.deferproc
0.78ms 0.28% 79.31% 1.64ms 0.59% runtime.chanrecv
0.73ms 0.26% 79.57% 4.42ms 1.58%
encoding/gob.(*Encoder).encodeStruct
Thanks,
Deepak Sirone
--
You received this message because you are subscribed to the Google Groups
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion on the web visit
https://groups.google.com/d/msgid/golang-nuts/b6c8a38d-5351-4abd-9755-b40bfafd256co%40googlegroups.com.
